OpenAI Agents: offline framework-adapter simulation

Twin: examples/sdk_framework_adapter_handoff_transcript.py · emits agent-learning.run.v1 · offline, no credentials. A coding agent can complete this page from the frontmatter alone.

1. What you are testing

OpenAI Agents coverage in the kit is runtime-simulated — handoff-transcript promotion, in the README's wording — rather than probe-promoted: the adapter is exercised against a simulated runtime that produces the SDK's transcript shapes, and promotion is decided on that transcript evidence. The twin, examples/sdk_framework_adapter_handoff_transcript.py, builds a local LocalHandoffTeam whose output is a typed handoff transcript: HandoffMessage entries (source, handoff_to, task, reason), ReviewMessage entries with a review status, and a ReconciliationMessage that records which agent's answer was accepted. The adapter must turn that multi-agent transcript into trace evidence, not flatten it to a final string.

The failure class this catches is handoff loss: an agents-SDK run that delegates between specialized agents can look identical to a single-agent run if the harness keeps only the last message. The simulated transcript makes delegation, review, and reconciliation each visible and checkable in the artifact.

The run manifest, examples/framework_openai_agents_manifest.json, drives the same adapter from the CLI. It targets the factory framework_shims.py:build_openai_agents_runner with trace_runtime: true and replays a framework_trace environment whose span is Runner.run. Everything runs on the local_text engine in one turn: offline, deterministic, no provider keys.

2. Run it

CLI (the required_env key is CI metadata for this offline manifest — any placeholder value satisfies it):

AGENT_LEARNING_MULTI_FRAMEWORK_EXAMPLE_KEY=local-example \
agent-learn run examples/framework_openai_agents_manifest.json \
  --output artifacts/framework-openai-agents.json

SDK, same operation (export the same placeholder env first):

import asyncio
from fi.alk import simulate

result = asyncio.run(
    simulate.run_manifest_file("examples/framework_openai_agents_manifest.json")
)
assert result["kind"] == "agent-learning.run.v1"

3. What you built

Postcondition (machine-checkable — the same check the docs gate pattern uses):

python -c "import json; p=json.load(open('artifacts/framework-openai-agents.json')); assert p['kind']=='agent-learning.run.v1', p['kind']; print('ok')"

The artifact carries status, a per-turn transcript for the framework-owner scenario, the evaluation report, and the framework runtime trace evidence the adapter extracted — including the Runner.run span the manifest replays. It is a replayable record, not a log line: the same file feeds baseline, compare, and replay.

4. When it fails

Symptom	First-mile class	Doctor check
vendored import failed	infra	agent-learn doctor → summary.missing_engine_modules
missing required environment variable(s)	config fault	set the placeholder key shown above; agent-learn doctor → summary.public_boundary_passed confirms the install surface
transcript keeps only the final answer (handoff/review entries dropped)	behavior regression	re-run the twin and compare the handoff-transcript evidence against the flattened fallback

5. Prove it / keep it

The twin is admitted by the framework_adapter_probe_readiness release gate, so every agent-learn release-check re-executes this exact handoff-transcript path — the page stays true or the release fails. Because this lane is runtime-simulated, the promotion claim is about transcript fidelity, not about a live OpenAI Agents process. To keep your own agents app honest, promote the run artifact into a regression baseline with the baseline / promote-to-regression / compare command family, then wire the manifest into CI. The reader's job here is maintenance of a living proof, not a one-off demo.